Quick Answer
AI can repeat a mistake you've already corrected because a correction in one interaction isn't necessarily a permanent change to the underlying system or to every future task.
We learned to classify corrections instead of assuming they would automatically stick.
A mistake may require a one-time fix, a persistent project rule, better source material, a workflow change, or acceptance that the task needs verification every time.
That distinction reduced repeated failures much more effectively than repeatedly saying, "I already told you this."
The Frustrating Pattern
We experienced this repeatedly.
AI produces something wrong. We identify the problem. It corrects it. Later, a similar mistake returns.
From a human perspective, the reaction is obvious:
"We already fixed this."
But that assumes the earlier correction is governing the current output in the same way a person might remember a lesson.
That's not always a safe assumption.
Some Corrections Are Local
Imagine AI uses the wrong heading on one article.
We correct the heading. That may be all that's needed.
Turning every typo into a permanent project rule would create an enormous instruction set.
Recurring Mistakes Are Different
Now imagine the same problem occurs across ten outputs.
That's evidence.
Something larger may be wrong: the instruction is ambiguous, the relevant rule isn't available when needed, another requirement conflicts with it, the source material is inconsistent, or the workflow itself encourages the failure.
Now the correction belongs above the individual output.
This Is How Workflows Become Bloated
Our early response to recurring mistakes was often: add another rule.
That worked sometimes. It also created a second problem.
The more rules we accumulated, the more interactions existed among them. Eventually the system designed to prevent repeated mistakes could itself produce new mistakes.
That led us to a better question:
"What level does this correction belong at?"
Four Levels of Correction
Output correction — Fix this result.
Task correction — Clarify how this kind of task should be performed.
Project correction — Change the current source of truth.
Workflow correction — Redesign the process that keeps producing the failure.
Those are very different responses.
Don't Rely on 'You Should Remember'
For important requirements, we prefer making the current requirement explicit and available to the task rather than relying on the emotional logic of:
"But we talked about this before."
The goal is reliable production, not winning an argument with the model.
Current Prompting Guidance Supports Iteration
OpenAI's current prompting guidance recommends reviewing outputs and refining instructions based on what happened rather than expecting a perfect initial prompt.
Our addition would be:
When the same refinement keeps returning, investigate the system instead of performing the same correction forever.
Worthy Trying Verdict
Does one correction guarantee permanent compliance? No
Can recurring mistakes be reduced? Yes
Best response: Classify the failure before adding another rule
Human oversight: Medium
Would we keep correcting the same symptom indefinitely? No
The lesson: When AI repeats a corrected mistake, don't just correct it again. Determine whether the problem belongs to the output, instruction, project standard or workflow.